WO2023019427A1 - Procédé et appareil de recommandation basée sur un graphe - Google Patents

Procédé et appareil de recommandation basée sur un graphe Download PDF

Info

Publication number
WO2023019427A1
WO2023019427A1 PCT/CN2021/112996 CN2021112996W WO2023019427A1 WO 2023019427 A1 WO2023019427 A1 WO 2023019427A1 CN 2021112996 W CN2021112996 W CN 2021112996W WO 2023019427 A1 WO2023019427 A1 WO 2023019427A1
Authority
WO
WIPO (PCT)
Prior art keywords
nodes
item
user
node
graph
Prior art date
Application number
PCT/CN2021/112996
Other languages
English (en)
Inventor
Evgeny Kharlamov
Jie Tang
Zhen Yang
Ming Ding
Xu Zou
Original Assignee
Robert Bosch Gmbh
Tsinghua University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Robert Bosch Gmbh, Tsinghua University filed Critical Robert Bosch Gmbh
Priority to PCT/CN2021/112996 priority Critical patent/WO2023019427A1/fr
Priority to CN202180101577.XA priority patent/CN118020067A/zh
Publication of WO2023019427A1 publication Critical patent/WO2023019427A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/901—Indexing; Data structures therefor; Storage structures
    • G06F16/9024—Graphs; Linked lists
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning

Definitions

  • aspects of the present disclosure relate generally to artificial intelligence and/or data mining, and more particularly, to a method and an apparatus for graph-based recommendation.
  • Graph-based recommendation systems are blossoming recently, in which the number of nodes in a graph may be usually huge. Requiring all nodes in the graph to be present during training of the embeddings of the nodes may be impractical, especially for billion-scale datasets.
  • Massive network embedding works have investigated good criteria to sample positive node pairs, such as by random walk, the second-order proximity, community structure, etc.
  • the strategy for negative sampling may be relatively unexplored, especially in graph representation learning, or may exist a large room for improvement. Recent studies have demonstrated that a negative sampling distribution may have the same level of influence on optimization as a positive sampling distribution, and the best choice may vary largely on different datasets.
  • the core idea of graph-based recommendation systems is to maximizing the outputting log-likelihood for embedding vectors in the observed user-item interactions which the users do react positively towards the items (which may be referred to as positive node pairs that may be obtained by positive sampling, for example) , while minimizing that otherwise (e.g., for negative node pairs that may be obtained by negative sampling) .
  • positive node pairs which may be obtained by positive sampling, for example
  • negative node pairs e.g., for negative node pairs that may be obtained by negative sampling
  • there may exist a crucial challenge in the negative sampling that only positive items are observed in the user-item graphs, while the other items may be not exposed to users and users’ opinions toward them are unknown.
  • the number of unexposed items is usually huge, and determining a best choice of the negative sampling distribution to obtain negative items from the huge number of items even without users’ attitudes may be an intractable problem.
  • a method for processing a graph having a plurality of nodes and a plurality of edges among the plurality of nodes, wherein the plurality of nodes comprise one or more user nodes and one or more item nodes comprising: obtaining a subgraph of the graph for a given user node by traversing the graph from the given user node, dividing the subgraph into a plurality of regions, wherein the plurality of regions comprise a first region adjacent to the given user node, a second region distant to the given user node and an intermediate region between the first and second regions, and performing sampling in the intermediate region to obtain a set of negative item nodes for the given user node.
  • a method for providing embeddings of nodes in a graph having a plurality of nodes and a plurality of edges among the plurality of nodes, wherein the plurality of nodes comprise one or more user nodes and one or more item nodes
  • the method comprising: receiving a current user embedding of a user node in the graph from a user encoder and current item embeddings of item nodes in a neighborhood of the user node from an item encoder, respectively aggregating the current user embedding of the user node with one or more current item embeddings of one or more item nodes in the neighborhood along respective types of edges, to obtain respective aggregated user embeddings, and integrating each of the aggregated user embeddings to obtain an integrated user embedding to replace the current user embedding of the user node.
  • apparatus for processing a graph having a plurality of nodes and a plurality of edges among the plurality of nodes, wherein the plurality of nodes comprise one or more user nodes and one or more item nodes comprises a memory and at least one processor coupled to the memory.
  • the at least one processor is configured to obtain a subgraph of the graph for a given user node by traversing the graph from the given user node, divide the subgraph into a plurality of regions, wherein the plurality of regions comprise a first region adjacent to the given user node, a second region distant to the given user node and an intermediate region between the first and second regions, and perform sampling in the intermediate region to obtain a set of negative item nodes for the given user node.
  • a computer program product for processing a graph having a plurality of nodes and a plurality of edges among the plurality of nodes, wherein the plurality of nodes comprise one or more user nodes and one or more item nodes comprises processor executable computer code for obtaining a subgraph of the graph for a given user node by traversing the graph from the given user node, dividing the subgraph into a plurality of regions, wherein the plurality of regions comprise a first region adjacent to the given user node, a second region distant to the given user node and an intermediate region between the first and second regions, and performing sampling in the intermediate region to obtain a set of negative item nodes for the given user node.
  • a computer readable medium stores computer code for processing a graph having a plurality of nodes and a plurality of edges among the plurality of nodes, wherein the plurality of nodes comprise one or more user nodes and one or more item nodes.
  • the computer code when executed by a processor causes the processor to obtain a subgraph of the graph for a given user node by traversing the graph from the given user node, divide the subgraph into a plurality of regions, wherein the plurality of regions comprise a first region adjacent to the given user node, a second region distant to the given user node and an intermediate region between the first and second regions, and perform sampling in the intermediate region to obtain a set of negative item nodes for the given user node.
  • the quality of the sampled negatives may be improved without sacrificing efficiency.
  • FIG. 1 illustrates a schematic diagram of an example graph-based recommendation system.
  • FIG. 2A shows an example user-item graph.
  • FIG. 2B shows an example subgraph for a given user.
  • FIG. 3A shows another example user-item graph.
  • FIG. 3B shows another example subgraph for a given user.
  • FIG. 4 illustrates an exemplary flow of a method for negative sampling according to a three-region principle, in accordance with one or more aspects of the present disclosure.
  • FIG. 5 illustrates another exemplary flow of a method for negative sampling according to the three-region principle, in accordance with one or more aspects of the present disclosure.
  • FIG. 6 illustrates an exemplary flow of a method for embedding propagating process with exposed information, according to one or more aspects of the present disclosure.
  • FIG. 7 illustrates an exemplary framework for performance of the method for embedding propagating process with exposed information, according to one or more aspects of the present disclosure.
  • FIG. 8 illustrates an example of a hardware implementation for an apparatus according to one or more aspects of the present disclosure.
  • Recommendation systems evolving from collaborative filtering to graph-based models may facilitate web-scale applications and show a promising prospect.
  • User-item graphs in E-commerce platforms may be constructed by user’s behavior history, which may contain information about user interests.
  • the key point of graph-based recommendation systems may be to learn embeddings for users and items via sampling technique.
  • the negative sampling strategy may accelerate the training process and reduce computational complexity, making it possible and efficient to implement large-scale graph-based recommendation system.
  • the quality of negative items does affect the user/item embedding qualities and the effectiveness of recommendation task.
  • the negative items should reflect user’s negative preferences over items, meanwhile, the sampled negative items should have enough discrimination such that the recommendation system can learn better embeddings to distinguish positive and negative items.
  • a classical strategy is to sample negative items with uniform distribution, which is efficient but may affect the model’s convergency and even the recommendation performance.
  • An effective solution for this problem is to sample hard (also referred to as “difficult” ) items based on the current model, which may be a heuristic negative sampler that however suffers from low efficiency. Therefore, negative sampling in graph-based recommendation system may need to be improved from perspectives of both efficiency and effectiveness.
  • FIG. 1 illustrates a schematic diagram of an example graph-based recommendation system, which may comprise a graph-based item encoder 110, a graph-based user encoder 120, a user-item graph 100, a positive sampler 140, a negative sampler 150, and a loss function 130.
  • the graph-based item encoder 110 and user encoder 120 may be used to learn embeddings of items and users, respectively.
  • the positive sampler 140 and negative sampler 150 may be used to sample positive and negative items respectively for any given user.
  • the sampled positive and negative user-item pairs may compose the training data for graph-based recommendation learning with stochastic gradient descent (SGD) optimizer.
  • SGD stochastic gradient descent
  • the loss function 130 in the graph-based recommendation system may be summarized to maximize the log-likelihood function of observed user-item pairs (i.e., positive user-item pairs) and to minimize that of negative user-item pairs.
  • the loss function 130 that is widely used in the graph-based recommendation system may be classified into two categories: pointwise loss and pairwise loss.
  • Binary cross-entropy loss is used to optimize graph-based encoders, which belongs to one kind of pointwise loss.
  • the pairwise loss functions are defined on a triplet of (u, v, v n ) , where the ratio of positive and negative items must be kept at 1: 1.
  • hinge loss is a commonly-used pairwise loss function in recommendation, aiming to keep a pre-defined safe margin between positive and negative user-item pairs.
  • the recommendation system may recommend the top-K items for a queried user (such as, the headphone, shirt and keyboard in FIG. 1) .
  • each node of a user-item graph (such as, user-item graph 100) may be categorized into one type based at least on the distance of that node from the given user.
  • the method for negative sampling according to the three-region principle may comprise sampling in the intermediate region for negative items for the given user.
  • sampling in the intermediate region for negative items may avoid futile items for the embedding learning.
  • sampling in the intermediate region may facilitate a hard or difficult negative item sampling.
  • sampling negative items according to the three-region principle may provide a sampled negative set specific to each user.
  • each node of the user-item graph 100A may be traversed to generate a subgraph 200A of FIG. 2B for the given user 210.
  • the subgraph 200A may be obtained using layer-wise Breadth First Search (LBFS) .
  • LBFS layer-wise Breadth First Search
  • the subgraph 200A of the given user 210 may be divided into three regions. For example, items within k hop-hop neighborhood of the given user 210 may be regarded as adjacent items, items outside k hop-hop neighborhood of the given user 210 may be regarded as distant items, and items in k hop-hop neighborhood of the given user 210 may be regarded as intermediate items.
  • k may be set to 3, and items 2, 4, 6, 8 may be in the adjacent region, items 10 and 12 may be in the intermediate region, and items 14 and 16 may be in the distant region, for the given user 210, as shown in FIG. 2B.
  • the negative sampling may be reviewed from a perspective of bias and variance.
  • the optimal embedding for each user-item pair (u, v) may satisfy:
  • p n denotes a negative sampling distribution
  • p d denotes a positive sampling distribution.
  • Distant items from u may be negatively sampled less.
  • a distant item v may have a small p d (v
  • u) may lead to rapidly approach negative infinity, and performance of negative sampling on them may be futile.
  • adjacent items to u may not be sampled, which are usually regarded as positive items. These items are usually used for propagating feature information for the given user u in graph-based recommendation system, and performance of negative sampling on them may not be meaningful either.
  • Intermediate items to u may be sufficiently sampled. Intermediate items may contain hard or difficult items that can bring more information for model training.
  • a set of one or more negative items for the given user 210 may be sampled from items 10 and 12 in the intermediate region.
  • k may be set to other values (e.g., 5 or the like) depending on design preferences across a variety of applications.
  • other dividing methods to the subgraph may be possible, for example, the subgraph may be divided into an adjacent region comprising one or more hop neighborhoods of the given user node (such as, 1 hop neighborhood consist of items 2, 4, 6, 8 of FIG. 2B) , and an intermediate region comprising one or more hop neighborhoods of the given user node (such as, 3, 4, 5 hop neighborhoods consist of items 10, 12, 14, 16 and user 230 of FIG. 2B) .
  • Exposure information may contain abundant information about the users’ negative preferences.
  • the expose information may be utilized in negative sampling.
  • Exposed but non-interacted items (abbreviated referred to as exposed items herein) may be more negative compared with unexposed items.
  • a strategy of sampling negative items from exposed items may face sampling bias since the exposed items themselves may be heavily biased, resulting in suboptimal performance. For example, ignoring the exposed items by a user may be due to negative attitudes of the user towards the items, but may also be possible resulted from a quick browsing or the like without a subjective attitude of the user. To reduce the bias, negative items may not only be sampled from the exposed items, but also unexposed items.
  • FIG. 3A illustrates an example user-item graph 100B, wherein the solid edges may represent interacted relations between users 310, 320, 330, 340 and items 2, 4, 6, 8, 10, 12, 14, 15, 16, the dotted edges may represent exposed relations that although items 2, 3, 5, 10, 13, 16 are exposed to users but users do not interact with them, and the user-item graph 100B may be an example of the user-item graph 100 of FIG. 1.
  • a subgraph 200B may be obtained for the given user 310.
  • a candidate set may be established for negative sampling for each user of the graph (e.g., graphs 100A and 100B) .
  • exposed items to the given user 310 may reflect more of the user’s negative preferences in some cases, and may be also included in the candidate set for negative sampling.
  • exposed items 3, 5, 10 to the given user 310 in the adjacent region may also be included in the candidate set, such that the candidate set can comprise both exposed and unexposed items to the given user 310.
  • the candidate set specific to the given user 310 comprises exposed items 3, 5, 10 and unexposed items 14, 16, 12 and 13.
  • FIG. 4 illustrates an exemplary flow of the method 400 for negative sampling according to the three-region principle, in accordance with one or more aspects of the present disclosure.
  • the method 400 may be performed in the example graph-based recommendation system of FIG. 1 or other systems.
  • a subgraph of the graph e.g., user-item graph 100, 100A or 100B
  • the subgraph may be divided into a plurality of regions, wherein the plurality of regions may comprise a first region adjacent to the given user node, a second region distant to the given user node and an intermediate region between the first and second regions.
  • sampling may be performed in the intermediate region to obtain a set of negative item nodes for the given user node.
  • the graph may be the user-item graph 100, 100A, 100B of FIG. 1, FIG. 2A and FIG. 3A, and have a plurality of nodes and a plurality of edges among the plurality of nodes, wherein the plurality of nodes may comprise one or more user nodes and one or more item nodes.
  • the plurality of edges may comprise an interacted edge between a user node and an item node that the user node interacted with (such as solid edges of FIG. 2A and FIG. 3A) .
  • the plurality of edges may also comprise an exposed edge between a user node and an item node that is exposed to the user node but the user node does not interact with (such as dotted edges of FIG. 3A) .
  • a candidate set may be established for the given user for negative sampling.
  • one or more items in the intermediate region of the given user may be included in the candidate set.
  • one or more items in the intermediate region and also exposed items in the adjacent region of the given user may be added to the candidate set for negative sampling.
  • hard or difficult negative samples may refer to those with a high probability of being positive according to the model, which are hard for learning.
  • hard or difficult negative items may be positive-likeness items that may have a relative high preference score than that of other negative items, and one or more items in the established candidate set with a high user preference sore may be sampled as hard negatives.
  • the user preference score of an item may be based on an inner production between embedding of the given user and embedding of that item.
  • the user-item graph when incorporating with the exposure information to enhance the performance of graph-based recommendation, may comprise interacted edges and exposed edges in the same user-item graph.
  • an adaptively sampling may be performed to adaptively mine real negative items and hard negative items in the candidate set according to the current model. For example, item nodes in the candidate set having a high preference score as hard negatives and item nodes in the candidate set having an exposed edge with the given user node as real negatives may be adaptively sampled, according to a factor for amplifying probabilities of real negatives and a distribution as a function of preference score of an item node in the candidate set derived from current embeddings generated by the encoder.
  • the factor may equal the number of item nodes in the candidate set having an exposed edge with the given user node for an item node in the candidate set having an exposed edge with the given user node, and the factor may equal one for an item mode in the candidate set that does not have an exposed edge with the given user node.
  • the factor may amplify probabilities of real negatives by the number of exposed items, more exposed items may be sampled as real negatives when there are more exposed items in the candidate set. In a case of more exposed items existed in the candidate set, these exposed items may be more likely to reflect the user’s real negative attitudes.
  • these exposed items themselves may be biased, and sampling a lot of these biased exposed items may result in suboptimal performance of the recommendation system.
  • negative items may be sampled from the exposed items and the unexposed items adaptively according to the amplifying factor.
  • FIG. 5 illustrates another exemplary flow of the method 500 for negative sampling according to the three-region principle, in accordance with one or more aspects of the present disclosure.
  • the method 500 may be performed in the example graph-based recommendation system of FIG. 1 or other systems.
  • a subgraph of the graph e.g., user-item graph 100, 100A or 100B
  • a given user node may be obtained by traversing the graph from the given user node (e.g., subgraph 200A or 200B for the given user 210 or 310 of FIG. 2B and FIG. 3B) .
  • the subgraph may be divided into a plurality of regions, wherein the plurality of regions may comprise a first region adjacent to the given user node, a second region distant to the given user node and an intermediate region between the first and second regions.
  • a candidate set may be established to include one or more item nodes in the intermediate region.
  • sampling may be performed on the candidate set to obtain the set of negative item nodes for the given user node. For example, the sampling at block 540 may be performed by sampling item nodes in the candidate set having a high preference score as hard negatives.
  • the method 500 may also comprise an optional block 535, where one or more item nodes in the first region having exposed edges with the given user node may be also added into the candidate set.
  • the sampling at block 540 may be performed by adaptively sampling between item nodes in the candidate set having a high preference score as hard negatives and item nodes in the candidate set having an exposed edge with the given user node as real negatives, according to a factor for amplifying probabilities of real negatives and a distribution as a function of preference score of an item node in the candidate set derived from current embeddings generated by the encoder.
  • the hardest negative item may be approximated via sampling item with the highest preference score between the given user and candidate negative items.
  • the distribution may be balanced by a similarity between an item node that the given user node interacts with (i.e., a positive item to the given user) and an item node in the candidate set, to avoid sampling an item node in the candidate set with a highest preference score.
  • the positive user-item pairs may provide more information to discover high-quality negative items, and such discovered negatives are closer to the positives, so as to offer meaningful gradients for training.
  • the candidate set may be established at block 530 (in conjunction with optional block 535) via the following Algorithm 1.
  • sampling at block 540 may be performed via the following distribution:
  • ⁇ ( ⁇ ) is the sigmoid function
  • v n denotes negative item
  • (u, v) denotes the positive pair
  • equation (2) inner productions may be conducted between user u and items in the candidate set C u based on the current embeddings outputted by the encoders (such as, graph-based user encoder 120 and graph-based item encoder 110 of FIG. 1) to represent respective preference scores of the items in the candidate set C u to the user u. Additionally, the similarity between the positive item v and items in the candidate set C u may be used to balance the equation (2) , to avoid the hardest negative items, for example. By using the distribution of equation (2) , a set of negative items may be sampled for each positive user-item pair (u, v) .
  • the coefficient of 1 in equation (2) may be replaced by an amplifying factor ⁇ , which may be formulated as:
  • Algorithm 2 shows the learning process with the present method (e.g., method 400 and/or method 500) for negative sampling according to the three-region principle.
  • exposed information may be incorporated into embeddings propagating and refining processes. Additionally, an augmented hinge loss may be used to increase the diversity of negative samples.
  • FIG. 6 illustrates an exemplary flow of method 600 for embedding propagating process with exposed information, according to one or more aspects of the present disclosure.
  • the method 600 may be performed in the graph-based recommendation system of FIG. 1 and other systems, and may be used with the method 400 and/or method 500 described above with reference to FIG. 4 and FIG. 5.
  • a current user embedding of a user node in the graph e.g., user-item graph 100, 100A or 100B of FIG. 1, FIG. 2A and FIG. 3A
  • a user encoder e.g., the graph-based user encoder 120 of FIG.
  • the current user embedding of the user node may be aggregated with one or more current item embeddings of one or more item nodes in the neighborhood along respective types of edges respectively, to obtain respective aggregated user embeddings.
  • the respective types of edges may comprise a first type of interacted edge between a user node and an item node that the user node interacted with, and a second type of exposed edge between a user node and an item node that is exposed to the user node but the user node does not interact with.
  • each of the aggregated user embeddings may be integrated to obtain an integrated user embedding to replace the current user embedding of the user node.
  • FIG. 7 illustrates an exemplary framework for performance of the method 600 for embedding propagating process with exposed information, according to one or more aspects of the present disclosure.
  • Block 120 may be an example of the graph-based user encoder 120 of FIG. 1
  • block 110 may be an example of the graph-based item encoder 110 of FIG. 1.
  • block 710-1 and block 710-2 e.g., aggregator
  • user embedding may be aggregated with items’ information over different edge types r i .
  • the following function may be used:
  • the aggregation function Aggre user may capture user’s preference on items and model user embeddings by aggregating over item embeddings.
  • the aggregator at block 710-1 may aggregate a user embedding with the user’s exposed items along an edge type of exposed to obtain the aggregated user embedding 711-1
  • the aggregator at block 710-2 may aggregate the user embedding with the user’s positive items along a different edge type of interacted to obtain the aggregated user embedding 711-2.
  • the aggregation operation at block 710-1 and block 710-2 may comprise one or more of mean-pooling that treats the neighbors equally, attention-pooling that differentiates the importance of neighbors with attention mechanism, degree-normalization that assigns weights to neighbors based on the graph structure, or central-filtering that uses the central node to filter the neighbors' information.
  • user embedding vector in l-th layer may be modeled as a combination of its neighbors’ item embeddings and its own attributes.
  • multiple-aggregation layer may be utilized to obtain the final user embedding by integrating the embeddings of all layers with weighted-pooling or concatenation operation.
  • the sampler operation may comprise one or more of identical operator that outputs the full neighborhood, sampling a fixed number of item nodes (e.g., allowing both up-sampling and down-sampling) , sampling neighbors in each convolutional layer, or adaptive layer-wise sampling.
  • item aggregation function Aggre item may be used to learn item embeddings. For example, for each item v, item embedding may be aggregated with exposed users at block 710-3 to obtain the aggregated item embedding 711-3 and positive users at block 710-4 to obtain the aggregated item embedding 711-4.
  • attention mechanism e.g., at block 720-1 and block 720-2
  • attention mechanism may be employed to assign weights to different types of edges (such as, the dotted edges representing exposed relation, and the solid edges representing positive or interacted relation of FIG. 7) .
  • edges such as, the dotted edges representing exposed relation, and the solid edges representing positive or interacted relation of FIG. 7 .
  • the l-th level user embedding vector may be calculated by self-attention mechanism across embeddings of multiple edge types, as demonstrated in the following exemplary formulation:
  • the attention weights may be obtained with a fully-connected neural network formulated as:
  • w 1 and w 2 denote trainable parameters for attention layer to learn the influence of edge type.
  • the integrated user embedding 721-1 and the integrated item embedding 721-2 may be provided to a hinge loss 730.
  • the current user embedding of the user node replaced by the integrated user embedding at block 650 and the current item embedding of the item node replaced by the integrated item embedding may be provided to a hinge loss for a next iteration process of optimization in the embedding learning.
  • the hinge loss is required to sample the same number of positive and negative pairs. It may be an insufficient exploration for negative sampling that samples only one negative item for each positive item. Empirically, sampling more negative items for each positive item may yield better performance.
  • the hinge loss 730 may comprise an augmented hinge loss that may be formulated as:
  • m different negative items v n s may be sampled for each positive item v to optimize the hinge loss between m pairs of v and v n s.
  • FIG. 7 may comprise one or more additional components other than the blocks illustrated in FIG. 7, depending on a design preference and/or a specific implementation, without departure of the present disclosure.
  • FIG. 8 illustrates an example of a hardware implementation for an apparatus 800 according to one or more aspects of the present disclosure.
  • the apparatus 800 for processing a user-item graph may comprise a memory 810 and at least one processor 820.
  • the processor 820 may be coupled to the memory 810 and configured to perform the method 400, the method 500, and method 600 described above with reference to FIGs. 4, 5 and 6.
  • the processor 820 may be a general-purpose processor, or may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.
  • the memory 810 may store the input data, output data, data generated by processor 820, and/or instructions executed by processor 820.
  • a computer program product for processing a user-item graph may comprise processor executable computer code for performing the method 400, the method 500, and method 600 described above with reference to FIGs. 4, 5 and 6.
  • a computer readable medium may store computer code for processing a user-item graph, the computer code when executed by a processor may cause the processor to perform the method 400, the method 500, and method 600 described above with reference to FIGs. 4, 5 and 6.
  • Computer-readable media includes both non-transitory computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. Any connection may be properly termed as a computer-readable medium. Other embodiments and implementations are within the scope of the disclosure.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Mathematical Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Pure & Applied Mathematics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Algebra (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

Procédé de traitement d'un graphe ayant une pluralité de nœuds et une pluralité de bords parmi la pluralité de nœuds, la pluralité de nœuds comprenant un ou plusieurs nœuds d'utilisateur et un ou plusieurs nœuds d'élément. Le procédé consiste à obtenir un sous-graphe du graphe pour un nœud d'utilisateur donné en traversant le graphe à partir du nœud d'utilisateur donné, à diviser le sous-graphe en une pluralité de régions, la pluralité de régions comprenant une première région adjacente au nœud d'utilisateur donné, une seconde région distante du nœud d'utilisateur donné et une région intermédiaire entre les première et seconde régions, et à effectuer un échantillonnage dans la région intermédiaire pour obtenir un ensemble de nœuds d'élément négatifs pour le nœud d'utilisateur donné.
PCT/CN2021/112996 2021-08-17 2021-08-17 Procédé et appareil de recommandation basée sur un graphe WO2023019427A1 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2021/112996 WO2023019427A1 (fr) 2021-08-17 2021-08-17 Procédé et appareil de recommandation basée sur un graphe
CN202180101577.XA CN118020067A (zh) 2021-08-17 2021-08-17 用于基于图的推荐的方法和装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/112996 WO2023019427A1 (fr) 2021-08-17 2021-08-17 Procédé et appareil de recommandation basée sur un graphe

Publications (1)

Publication Number Publication Date
WO2023019427A1 true WO2023019427A1 (fr) 2023-02-23

Family

ID=85239970

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/112996 WO2023019427A1 (fr) 2021-08-17 2021-08-17 Procédé et appareil de recommandation basée sur un graphe

Country Status (2)

Country Link
CN (1) CN118020067A (fr)
WO (1) WO2023019427A1 (fr)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8689108B1 (en) * 2013-09-24 2014-04-01 Palantir Technologies, Inc. Presentation and analysis of user interaction data
US20140095365A1 (en) * 2012-09-28 2014-04-03 Trading Technologies International, Inc. Directed Graph-Based Strategy Trading Tool
WO2014146265A1 (fr) * 2013-03-20 2014-09-25 Nokia Corporation Procédé et appareil pour des recommandations de ressources personnalisées
CN112732936A (zh) * 2021-01-11 2021-04-30 电子科技大学 一种基于知识图谱和用户微观行为的广电节目推荐方法
CN112966165A (zh) * 2021-02-03 2021-06-15 北京大学 基于图神经网络的交互式社区搜索方法及装置
US20210248449A1 (en) * 2020-02-12 2021-08-12 Jianing Sun Recommender system using bayesian graph convolution networks

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140095365A1 (en) * 2012-09-28 2014-04-03 Trading Technologies International, Inc. Directed Graph-Based Strategy Trading Tool
WO2014146265A1 (fr) * 2013-03-20 2014-09-25 Nokia Corporation Procédé et appareil pour des recommandations de ressources personnalisées
US8689108B1 (en) * 2013-09-24 2014-04-01 Palantir Technologies, Inc. Presentation and analysis of user interaction data
US20210248449A1 (en) * 2020-02-12 2021-08-12 Jianing Sun Recommender system using bayesian graph convolution networks
CN112732936A (zh) * 2021-01-11 2021-04-30 电子科技大学 一种基于知识图谱和用户微观行为的广电节目推荐方法
CN112966165A (zh) * 2021-02-03 2021-06-15 北京大学 基于图神经网络的交互式社区搜索方法及装置

Also Published As

Publication number Publication date
CN118020067A (zh) 2024-05-10

Similar Documents

Publication Publication Date Title
CN112989064B (zh) 一种聚合知识图神经网络和自适应注意力的推荐方法
CN110555050B (zh) 一种基于元路径的异构网络节点表示学习方法
Ning et al. Constrained multi-objective optimization using constrained non-dominated sorting combined with an improved hybrid multi-objective evolutionary algorithm
WO2022166115A1 (fr) Système de recommandation à seuils adaptatifs pour sélection de voisinage
CN110175286B (zh) 结合成对优化和矩阵分解的产品推荐方法及系统
CN116685985A (zh) 具有多样化反馈的联合学习系统与方法
CN112364242B (zh) 针对上下文感知型的图卷积推荐系统
WO2023045149A1 (fr) Procédé et appareil de fusion d'images, dispositif électronique et support de stockage
CN115271099A (zh) 一种支持异构模型的自适应个性化联邦学习方法
CN113918834B (zh) 融合社交关系的图卷积协同过滤推荐方法
Garg et al. Direct federated neural architecture search
US20230027427A1 (en) Memory-augmented graph convolutional neural networks
CN116340646A (zh) 一种基于超图基序优化多元用户表示的推荐方法
CN111353534B (zh) 一种基于自适应分数阶梯度的图数据类别预测方法
Azabou et al. Half-Hop: A graph upsampling approach for slowing down message passing
WO2022166125A1 (fr) Système de recommandation comprenant une perte de classement de bayes personnalisée pondérée adaptative
CN111738514B (zh) 利用局部距离和节点秩优化函数的社交网络社区发现方法
WO2023019427A1 (fr) Procédé et appareil de recommandation basée sur un graphe
CN111738516A (zh) 通过局部距离和节点秩优化函数的社交网络社区发现系统
CN114819196B (zh) 基于噪音蒸馏的联邦学习系统及方法
Thirunavukarasu et al. People to people recommendation using coupled nonnegative boolean matrix factorization
CN115757384A (zh) 一种基于大数据的政务数据处理方法
CN114332472A (zh) 一种基于图神经网络的数据处理方法及装置
Yang et al. A dynamic global backbone updating for communication-efficient personalised federated learning
Yang et al. A novel pso-de co-evolutionary algorithm based on decomposition framework

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21953679

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE